사내 AI 에이전트 온보딩 가이드 만드는 법

Let’s say one day you are tasked with guiding AI agent onboarding within the company. It seems simple at first. I think all you need to do is compile a list of approved tools, gather good examples of prompts, and add security precautions.

But in reality, another problem soon arises.

1This automation requires administrator privileges.2Where do I get an API key?3Can I include the original text of the customer inquiry?4Can operation logs be read by AI?5Can I immediately distribute code created by AI?6Which team budget covers expenses?

When this question starts to come up, onboarding is no longer just “how-to” training. It becomes the operational design that determines which tasks the organization will entrust to AI, which data will be blocked, and which results will be verified by humans.

An AI agent onboarding guide shouldn't just be a collection of prompts. The key is to answer the following questions:

1What tasks can be entrusted to AI agents?2What data should not be included?3Which practices must be approved?4How will humans verify AI results?5How will you change your team's workflow?

Key takeaways

• Your in-house AI agent onboarding guide should be a task transformation playbook, not a how-to document.
• AI agents can handle more tools and permissions than regular chatbots, so permissions/data/cost/verification criteria must be designed together.
• The introduction flow can be divided intoHarbor -> Spark -> Harness -> Route. First create formal operating boundaries, demonstrate small success stories, add secure guardrails, and then operate with continuously updated standards.
• It is better to separate onboarding documents into three sets:Quick Start Guide,Role-based Playbook, andGovernance & Safety Guide.
• Operational metrics should be viewed in terms of reduced repetitive work, reusable workflows, approved automation rates, and quality improvements rather than number of accounts or prompts.

Why is AI agent onboarding necessary?

This is how most AI tools start when introduced for the first time.

1Try ChatGPT.2Try Claude.3Try using Copilot.4Try using Cursor.5Try Claude Code.

This approach alone will not increase overall organization utilization. People want to know how to use a tool in their work rather than the name of the tool.

Developers are interested in PR reviews and test code generation. QA is concerned with test cases and reproducibility. Planners are interested in organizing requirements and drafting policy documents, while Operations/CS is interested in triaging customer inquiries and drafting responses.

Microsoft 2026 Work Trend Index explains that as AI and agents take over execution, the role of humans in directing and judging work increases. The interpretation in this article is clear. Internal onboarding should not be a document telling individuals to “try AI well,” but rather an operational document for the organization to absorb AI safely and repeatably.

Differences between AI agents and regular chatbots

This difference should be explained briefly and clearly at the beginning of onboarding.

Regular chatbots mainly answer questions. An AI agent receives a goal, finds the necessary information, calls a tool, and goes through several steps to produce a result.

OpenAI Agents SDK Documentation describes agents as LLM execution units that can be composed of runtime elements such as instructions, tools, handoffs, guardrails, and Structured Outputs. [OpenAI's agent construction guide] (https://openai.com/business/guides-and-resources/a-practical-guide-to-building-ai-agents/) also explains that agents are suitable for workflows with complex decisions, rules that are difficult to maintain, and a lot of unstructured data.

Your onboarding guide should include the following sentences:

1AI agents are action tools, not answer tools.2Therefore, as execution authority grows, verification standards and guardrails must also grow.

AX introduction methodology: Harbor, Spark, Harness, Route

The introduction of AX within the company is not a training project but an organizational operating system conversion project. This article divides it into four steps.

Harbor: Management creates the harbor

Organizational transformation will not occur if AI use remains at the level of individual stealth. First, the company must make it official.

The purpose of Harbor is not control. The idea is to bring the state of personal use within the boundaries of official operation.

Spark: creates longing

To get people to use AI, you don't need to explain the LLM principles first. Even stronger is the actual example of the team next door.

1A person on the next team used AI to reduce work that used to take 2 hours to 20 minutes each week.

Atlassian AI Use Case Demo Playbook suggests concrete use case demos as one way to encourage AI adoption. Live demos, especially ones that solve real-world problems, make colleagues think, "I could write like that."

Your onboarding guide should have more Before/After examples than tooltips.

Harness: Creating guardrails that accelerate freedom

As AI usage begins to take off, permission requests will increase. At this time, if you block everything, the spread will die, but if you open everything, an accident will occur.

So what you need is a Harness.

Harness is not a control that prevents the use of AI, but a guardrail that safely allows more use. NIST AI RMF Core divides AI risk management into Govern, Map, Measure, and Manage functions. In-house AI agents should be viewed in the same way.

Route: Standards keep changing

AI tools and models change rapidly. Today's best practice may be outdated next month. That’s why your onboarding guide should be a living document, not a finished document.

The point of the Route phase is not to fix standards, but to create a loop in which successes and failures are continually reflected.

1Collect new use cases2to risk classification3to be reflected in each team’s practice4to failure cases and guardrail updates5to be reflected in the following onboarding materials

The onboarding guide is divided into three sets.

If you put everything in one document, no one will read it to the end. It's best to split your internal onboarding guide into three documents:

Quick Start Guide

The success criterion for an introductory document is not the amount of explanation, but the first work product.

1target:2First-time user within 30 minutes3It makes you handle one of your tasks with an AI agent.

Role-based Playbook

Role-based Playbook must be divided by job category. Even with the same AI tool, the first success experience is different for each job.

Governance & Safety Guide

This document doesn't need to be long. But it shouldn't be ambiguous.

OpenAI Enterprise Privacy explains the ownership and control of business data in enterprise products and API Platform, whether model learning is used, and access control within the organization. However, in-house guides should not rely solely on specific vendor policies. Data processing standards for each tool actually used by the company must be organized in a separate table.

Design your first success experience

The most important thing about onboarding is experiencing your first success. If you try to automate something grandiose from the beginning, it will fail. It requires small exercises that can be completed in 15 to 30 minutes.

The important thing here is that the subject of the practice should be “Solving one of my tasks,” not “Experience AI functions.”

Use cases are created with cards for each occupation.

Good use case cards show workflow rather than prompts. Cards for developers can be created like this:

1# Use case card: Summary of PR changes2 3## Target4Developer, Tech Lead5 6## Conventional method7Check PR diff to Read changed files to Write review points to Organize items needed for testing8 9## AI application method10Enter PR diffs and related issues to Change Summary to Risk Points to Generate Test Recommendations11 12## Input data13- PR diff14- Related Jira/GitHub Issues15- Some changed files16- Existing test code17 18## Output format19- 5 line summary20- Major changes21- Danger Points22- Items required for testing23- Review Questions24 25## caution26- Check that secret keys, tokens, and customer information are not included in the diff.27- Do not approve AI results as is28- The final merge decision is made by a human.

For QA cards, the input and verification criteria are different.

1# Use case card: Requirements-based test case creation2 3## Target4QA, planner, developer5 6## Conventional method7Reading requirements documents to organizing normal flow to manually deriving exception cases8 9## AI application method10Enter requirements and screen policies to create normal/exception/boundary test drafts11 12## Output format13- Test purpose14- Prerequisites15- test data16- procedure17- Expected results18- Priority19 20## Verification criteria21- Are there any cases that conflict with business policies?22- Were any security/permission related cases left out?23- Is the procedure actually executable by humans?

AI agent use is classified into three stages:

Summarizing meeting minutes and automating access to operational databases have completely different risks. All uses of AI should not be managed at the same level.

Google Cloud's Generative AI Security Guide covers generative AI workloads from the perspectives of infrastructure, data management, and application control. Google Cloud Production-Ready AI Learning Path also divides agents into development, security, deployment, evaluation, and production patterns.

Your in-house guides should be the same way. The starting point is not “Use AI,” but “First classify what level of work it is.”

Security/data/permission standards should be written on 1 page.

If your security document is too long, no one will read it. Your onboarding guide should have a one-page “what you can/don’t include” table.

Data prohibited from entering

First, prohibited data must be displayed directly in a checklist.

1What you should not input into an AI agent2 3[ ] Customer personal information4[ ] Resident registration number, account number, card number5[ ] Authentication token, API Key, Secret Key6[ ] Original text of non-disclosure agreement7[ ] Documents from other teams that you do not have access to8[ ] Operation DB original data9[ ] Security vulnerability details10[ ] Source code prohibited from being disclosed externally

Conditionally accepted data

Anonymization and masking standards must be written down so that users can immediately judge them.

1Conditionally available2 3[ ] Anonymized logs4[ ] Masked customer inquiry5[ ] Part of code that can be made public6[ ] Summary of internal policies7[ ] Dummy data for testing8[ ] Sample API response

Permission request criteria

Once AI agents start calling tools, permission management becomes important.

Operates as a 4-week onboarding program

If you only distribute documents, the usage rate will not increase. Onboarding should be a program, not a document.

The advantages of this approach are clear.

• People actually use it.
• There are success stories.
• Failure cases are also reflected in the documentation.
• The guide becomes a living document.

Practical template

At the end of your onboarding guide, you should include a template that you can copy and use right away.

AI agent request template

1# AI agent request template2 3## 1. Goal4What needs to be done?5 6## 2. Background7Why is this necessary?8 9## 3. Input data10What documents, logs, code, issues, and policies should AI refer to?11 12## 4. Constraints13What should not be done, what rules should be followed, and what are the formalities?14 15## 5. Output format16Will you receive it as a table, Markdown, JSON, checklist, code, or email draft?17 18## 6. Verification criteria19What must a person confirm to be considered complete?

The quality of the request significantly changes the quality of the result.

AX Automation Request Form

For work above Level 2, it is better to receive a request form.

1# AX Automation Request2 3## 1. Problem you are trying to solve4-What tasks are currently repeated?5- How long does it take to process once?6- How many times per month does it occur?7 8## 2. Current work flow9Before:101.112.123.13 14## 3. Expected flow after applying AI15After:161.172.183.19 20## 4. Data used21- Public data:22- Internal general data:23- Sensitive data:24- Customer/Personal Information:25- Whether source code/private key is included:26 27## 5. Required permissions28- File access:29- API access:30- DB access:31- External service integration:32- Requires administrator privileges:33 34## 6. Result35- document36- code37- Report38- alarm39- Web service40- Auto-run script41 42## 7. Risk classification43- Level 1: Personal productivity44- Level 2: Team work automation45- Level 3: System/data linkage46 47## 8. Verification criteria48- Items for humans to check:49- How to revert in case of failure:50- Logs or evidence:

Checklist before personal use

1# Checklist before using AI agent2 3## Input data4[ ] Does it contain customer personal information?5[ ] Are the API Key, Secret, and Token included?6[ ] Does it contain any contracts/policies/source code that cannot be disclosed externally?7[ ] Has it been anonymized or masked where necessary?8 9## Scope of work10[ ] Is the task to be entrusted to AI clear?11[ ] Are the items for human final judgment separate?12[ ] Did you specify the output format?13[ ] Is it possible to turn back in case of failure?14 15## verification16[ ] Have you confirmed the facts created by AI?17[ ] Has the AI-generated code been tested?18[ ] Have you checked for security/permissions impacts?19[ ] Are there excessive costs?20 21## share22[ ] Can this use case be reused by my team?23[ ] Was it organized into a workflow rather than a prompt?24[ ] Are success/failure stories posted on internal channels?

Operational indicators are viewed as changes in business rather than usage.

To encourage active use of AI agents, you must also have good metrics. Many organizations first look at usage.

1Number of accounts issued2Number of prompts per month3Number of users accessing AI tools

However, it is difficult to know the actual effect from this alone. A more important indicator is the change in workflow.

The goal of your onboarding guide should not be to increase usage, but to answer the questions below:

1How has this team changed their workflow with AI agents?2Is the change safe?3Is it repeatable?4Can other teams also use it?

Order of operation starting immediately

There is no need to create a huge AX organization from scratch. You can start small.

11. Collect three tasks that each team repeats every week.22. Classify each task into Level 1/2/3.33. Create a 30-minute practice with one Level 1 task.44. Leave before/after cases on a card.55. Include prohibited data and permission standards on one page.66. After 4 weeks, update the playbook to reflect actual cases and failure cases.

AX is not a project that uses a lot of AI. A project to redesign repeatable workflows so people can spend their time on more important issues.

Not a collection of prompts but a business playbook, not a tool introduction but a first success experience, and risk-based guardrails rather than unconditional permission. If you stick to these three things, onboarding in-house AI agents goes beyond training material and becomes an operating system that changes the way your organization runs.

summary card

References

• OpenAI Agents SDK: Agents
• OpenAI: A practical guide to building agents
• OpenAI Enterprise Privacy
• Microsoft 2026 Work Trend Index
• NIST AI RMF Core
• Atlassian AI Use Case Demo Playbook
• Google Cloud security best practices for generative AI workloads
• Google Cloud Production-Ready AI Learning Path